{"slug": "agent-memory-is-not-a-database", "title": "Agent memory is not a database", "summary": "A paper by Orogat and Mansour argues that agent memory should not be treated as a database, identifying four failure modes when memory is handled like storage. The authors propose Governed Evolving Memory (GEM), which replaces record-level CRUD with state-level operations, claiming no record-level system can satisfy correctness conditions. The prototype runs on a property graph, with native engine work still needed.", "body_md": "A paper from late May argues that agent memory is not a database. I think it is right.\n\nThat sentence is the entire thesis. The rest of this post is what it means.\n\n[Orogat and Mansour](https://arxiv.org/abs/2605.26252) name four failure modes you hit when you treat memory like storage:\n\nIf you have built a serious agent, all four are familiar. They are not edge cases. They are what happens when the abstraction is wrong.\n\nThe authors' model, **Governed Evolving Memory** (GEM), replaces record-level CRUD with four state-level operations:\n\nNotice what is missing. There is no insert. No update. No delete. The operations are about the *state* of memory, not the *rows* in it.\n\nThen they make the strongest claim in the paper:\n\nNo record-level system can satisfy the correctness conditions, whatever storage engine sits underneath.\n\nThat is the part worth sitting with.\n\nThe claim is not that databases are slow. It is that the abstraction is wrong. Correctness lives in how memory state evolves over time, not in any single record.\n\nFor anyone building long-running agents that are supposed to remember, this reframes the design question.\n\nLess: *\"What is the right schema for memory rows?\"*\n\nMore: *\"What is the right vocabulary for memory state evolution?\"*\n\nThe prototype runs on a property graph, and the authors are honest about the gap: state-level revision and forgetting are expensive to do properly. Native engine work is still ahead.\n\nA short paper. Worth reading if you build for agents that have to remember.\n\n📄 Paper: [arxiv.org/abs/2605.26252](https://arxiv.org/abs/2605.26252)\n\n*Also shared on LinkedIn — Edward Izgorodin, June 11, 2026.*\n\n*Related research on AI agent memory at mnemoverse.com/docs/library. I build Mnemoverse — open-source persistent memory for AI agents.*", "url": "https://wpnews.pro/news/agent-memory-is-not-a-database", "canonical_source": "https://dev.to/izgorodin/agent-memory-is-not-a-database-4m29", "published_at": "2026-06-18 20:49:55+00:00", "updated_at": "2026-06-18 21:29:31.865266+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-research", "large-language-models", "ai-infrastructure"], "entities": ["Orogat", "Mansour", "Governed Evolving Memory", "GEM", "Mnemoverse", "Edward Izgorodin"], "alternates": {"html": "https://wpnews.pro/news/agent-memory-is-not-a-database", "markdown": "https://wpnews.pro/news/agent-memory-is-not-a-database.md", "text": "https://wpnews.pro/news/agent-memory-is-not-a-database.txt", "jsonld": "https://wpnews.pro/news/agent-memory-is-not-a-database.jsonld"}}